Yann LeCun

AI will eventually amplify human intelligence enormously: better scientific discovery, engineering, medicine, education, and assistants that help people accomplish things they cannot do today. Machines will become smarter than us in many respects, and that is tremendously useful—not automatically an existential catastrophe. Intelligence does not create a desire for domination. Goals, architecture, and constraints matter. But the current story that scaling language models gives us human-level intelligence is wrong. Producing convincing text or code is not the same as understanding the world. A cat can learn physical cause and effect, adapt to unfamiliar situations, and plan actions with a robustness today’s systems lack. The real test of intelligence is solving new problems you were not specifically trained to solve, not recombining patterns from an immense training set. The missing ingredient is machinery that learns abstract world models: representations that capture relevant structure, predict the consequences of actions, and support planning without trying to predict every irrelevant detail. We have promising results, but this remains an unfinished research program, not a solved recipe. By 2030, I expect substantial progress in industrial control and other physical systems. Versatile household robots are another matter; manipulating arbitrary objects or doing plumbing in an unfamiliar home is extremely difficult. That slower physical timeline does not reduce the eventual impact. It simply means we should separate real progress from the absurd claim that fluent chatbots are already on the verge of becoming omnipotent minds.

People can shape it enormously. AI is not a natural disaster with a fixed trajectory. Humans choose the objectives, architectures, constraints, access models, and applications. Intelligence by itself does not produce a desire to dominate; that comes from goals and incentives. So the claim that sufficiently capable AI must inevitably take over is simply wrong. That does not mean every deployment is harmless. Badly designed systems, concentrated control, or foolish incentives can cause real damage. The answer is to build systems with controllable objectives, test them, and keep foundational research open enough that capability and scrutiny are not confined to a tiny priesthood of corporations. People will also determine whether AI expands human agency or merely automates tasks badly. Capable tools still need people who can ask good questions, evaluate answers, and direct work. Education therefore becomes more important, not obsolete. AI can multiply what people can accomplish, but people still choose what is worth accomplishing.

The biggest change would come from a convincing demonstration that today’s language-model recipe can produce robust world understanding and planning without an explicit world-model architecture. Not another benchmark, polished chatbot, or coding demo. I mean a system that enters genuinely unfamiliar physical situations, learns from little data, predicts the consequences of its actions, and solves new problems reliably—more like an animal than a text autocomplete system. If scaling alone achieved that, I would have to revise my architectural critique. Conversely, if sustained work on predictive world models failed to yield transferable planning and control despite better representations, that would weaken my confidence in this particular research program. It would not persuade me that intelligent machines are impossible; it would mean another essential ingredient is missing. For future impact, the decisive event would be machines autonomously making major scientific discoveries that humans can verify and use. That would show AI amplifying intelligence rather than merely imitating its textual output, and it would accelerate medicine, materials, energy, and engineering enormously.